Intrafractional Tumor Motion Prediction Using Surrogates On Pancreatic Cine MRgRT
A Guta1*, J Shin2, B Lewis3, Z Ji4, J Kim5, T Kim6, (1) Washington University in St. Louis, St. Louis, MO, (2) Yonsei University College of Medicine, Seoul, ,KR, (3) Washington University School of Medicine in St. Louis, Saint Louis, MO, (4) Washington University St Louis, ,,(5) Yonsei University College of Medicine, Seoul, ,KR, (6) Washington University in St. Louis, Saint Louis, MO
Presentations
PO-GePV-M-32 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: X-ray image-guided radiotherapy (IGRT) has been a vital tool in radiation oncology to visualize and correct for inter- and intra- fractional variations during radiotherapy treatment. However, X-ray imaging lacks good soft tissue contrast, limiting direct tumor visualization. This work uses 2D cine MR images obtained on a 0.35T MRgRT system to create a pancreatic tumor position prediction model using surrogates visible on X-ray imaging.
Methods: A retrospective dataset of cine MRI images was collected from pancreatic tumor patients treated with 0.35T MRgRT at our institution, containing a total of 56 treatment fractions and 53,759 images. In-house software was used to extract tumor position with respect to diaphragm position from cine MR images. Diaphragm positions were determined using template-based auto-segmentation and tumor positions were extracted from the MRgRT systems.Diaphragm contours were manually evaluated and adjusted. Models were developed using ridge regression and feature reduction using principal component analysis (PCA). The intra-fractional prediction model was trained on each fraction using 70% of images and tested with the remaining 30%. The inter-fractional model was trained with all fractions except 10% of image in the target fraction, and tested with the remaining 90% of images in the target fraction. Models were evaluated with mean square error (MSE) and mean absolute error (MAE).
Results: The predicted pancreatic tumor position had an average MSE (mm2) of 0.34/0.98(X/Y),and average MAE(mm) of 0.42/0.68(X/Y),for the intra-and average MSE(mm2) of 0.90/2.1(X/Y), and average MAE(mm) of 0.66/1.00(X/Y) for the inter-fractional models, respectively. After PCA of contours the models predicted tumor position with an average MSE(mm2) of 0.37/1.1(X/Y) and average MAE(mm) of 0.45/0.74(X/Y),for the intra- and average MSE(mm2) of 2.0/6.2(X/Y),and average MAE(mm) of 0.99/1.8(X/Y) for the inter-fractional models,respectively.
Conclusion: Intra-and inter-fraction models were developed to predict pancreatic tumor position from diaphragm position, a surrogate visible on both MR and X-ray imaging.
ePosters
Keywords
Taxonomy
IM- Dataset Analysis/Biomathematics: Machine learning
Contact Email